If a W-cycle is used in the AMG the residual is not updated between the two subcycles, which leads to an inconsistency of the update and the residual. This MR adds this residual update. Furthermore it breaks after one iteration of the coarsest level, as the direct solver does not improve the solution further.
It also amends the amgtest that checks whether the convergence rate of the W-cycle is smaller than the convergence rate of the V-cycle. At least for the case used in the test that should hold true.
For the anisotropic example that is also used in the test the convergence rate of the GeneralizedPCGSolver is improved from 0.4 to 0.07 by this MR.
Nils-Arne Dreier authored